Add tests

tests/python/gpu/test_gluon_gpu.py

@@ -113,10 +113,13 @@ def test_gluon_ctc_consistency():
 def test_global_norm_clip_multi_device():
     x1 = mx.nd.ones((3,3), ctx=mx.gpu(0))
     x2 = mx.nd.ones((4,4), ctx=mx.cpu(0))
-    norm = gluon.utils.clip_global_norm([x1, x2], 1.0)
-    assert norm == 5.0
-    assert_almost_equal(x1.asnumpy(), np.ones((3,3))/5)
-    assert_almost_equal(x2.asnumpy(), np.ones((4,4))/5)
+    for check_isfinite in [True, False]:
+        for check_scale in [True, False]:
+            norm = gluon.utils.clip_global_norm([x1, x2], 1.0, check_isfinite=check_isfinite,
+                                                check_scale=check_scale)
+            assert norm == 5.0
+            assert_almost_equal(x1.asnumpy(), np.ones((3, 3)) / 5)
+            assert_almost_equal(x2.asnumpy(), np.ones((4, 4)) / 5)
 
 
 def _check_batchnorm_result(input, num_devices=1, cuda=False):

tests/python/unittest/test_gluon.py

@@ -735,10 +735,11 @@ def test_sequential_warning():
 @with_seed()
 def test_global_norm_clip():
     stypes = ['default', 'row_sparse']
-    def check_global_norm_clip(stype):
+    def check_global_norm_clip(stype, check_isfinite, check_scale):
         x1 = mx.nd.ones((3,3)).tostype(stype)
         x2 = mx.nd.ones((4,4)).tostype(stype)
-        norm = gluon.utils.clip_global_norm([x1, x2], 1.0)
+        norm = gluon.utils.clip_global_norm([x1, x2], 1.0, check_isfinite=check_isfinite,
+                                            check_scale=check_scale)
         assert norm == 5.0
         assert_almost_equal(x1.asnumpy(), np.ones((3,3))/5)
         assert_almost_equal(x2.asnumpy(), np.ones((4,4))/5)
@@ -750,7 +751,9 @@ def check_global_norm_clip(stype):
             assert len(w) == 1
 
     for stype in stypes:
-        check_global_norm_clip(stype)
+        for check_isfinite in [True, False]:
+            for check_scale in [True, False]:
+                check_global_norm_clip(stype, check_isfinite, check_scale)
 
 @with_seed()
 def test_embedding():